[arXiv]score: 0.15

SFT Beats CPT+SFT for French Medical QA in Cost-Efficiency

June 18, 2026

Across three model families and multiple sizes, CPT+SFT yields the highest MCQA scores for French medical QA, but gains over SFT alone are small and rarely statistically significant. For open-ended QA, CPT consistently improves overlap metrics. SFT emerges as the cost-effective default for medical domain adaptation.

HOW THIS AFFECTS YOU

●

researcherConcrete ablation across CPT, SFT, and CPT+SFT with statistical significance testing gives you a reliable reference point for medical domain adaptation strategy selection.

●

healthWorth watching because it quantifies when the added cost of continual pretraining is justified for clinical NLP tasks, particularly in non-English languages.

read original ↗arxiv.org

← back to feed